Implementation of MEBP learning circuitry with simple nonlinear synapse circuits
Myung‐Ryul Choi, Jin‐Sung Park · 1999
MEBP (Modified Error Back-Propagation) learning rule has been implemented using simple nonlinear synapse circuits. The simple nonlinear synapse circuit is suitable for implementation of artificial neural networks using standard CMOS technology since it requires large number of neurons. The learning circuitry consists of nonlinear synapse circuits, sigmoid circuits, and linear multipliers, whose output voltage is uniquely determined by any pair of learning input patterns. The proposed learning circuitry is applied for 2/spl times/2/spl times/1 and 2/spl times/3/spl times/1 multilayered feedforward neural network model. MEBP rule has been simulated successfully via C programmed software implementation. And its hardware implementations have been verified by using HSPICE circuit simulator. The proposed learning circuitry is very suitable for the future implementation of the large-scale neural networks or fuzzy processors including on-chip learning.